Loading activity…
@askalphaxiv
High fidelity research
Technology
In a 50-forward-pass test, Astra reportedly changed 0.016% of Qwen3.5-0.8B's weights, raising MATH-500 accuracy from 7.9% to 48.1%.
AI
Mistral's report drew praise for technical detail, while another reader questioned model surgery instead of starting over.
AI
The reported MATH-500 score rose from 7.9% to 48.1% with the same sampling settings.
AI
A 109M JEPA-Text model is claimed to align with vision models better than 72B LLMs.
XAI
A post also thanked participants in a fireside chat at the October 27 launch event.
AI
Its developer says Flint exceeds 125 tokens per second on an M4 Pro MacBook Pro at short contexts.
XAI
alphaXiv says the app pairs recent research with key ideas, figures and audio explanations from alphaXiv Briefs.
XAI
The announcement says H-JEPA builds on SIGReg/LeWM and describes semantic abstractions emerging across levels during training.
AI
A post describing the paper says its method uses the difference between early and final loops to adjust predictions without extra training.
AI
A post sharing @a1zhang’s blog describes an alternative to fitting an AI agent’s harness around a decoder-only Transformer: change the model’s input and output shape to fit the agent.
AI
askalphaxiv shares a paper describing “mode-hopping,” in which LLMs repeatedly shift between shallow pattern-matching and generalization.
AI
A post about the project claims a 65% score gain with the same compute on a 24-hour multi-repository agent-swarm task.
AI
The paper’s authors say MAttr uses gradient descent to find which parts of a neural network are responsible for a behavior. They report a benchmark result of 2.9× the runner-up.
AI
alphaXiv says its tool finds the shortest chain of coauthored papers connecting any two AI researchers on arXiv.
AI
alphaXiv says OpenResearch can turn any coding agent into a research agent that reviews literature, develops hypotheses, runs experiments and produces research artifacts.
AI
The reader says removing rollout groups lets completed runs enter a batch without waiting for slower runs in the same group.
AI
In its summary of “Recurrent Looped Transformer,” alphaXiv says the decoder carries state across every prompt and response token, while a causal encoder provides reusable global memory.
AI
Academic Mohit Iyyer retweets claim that machine translation remains unsolved.
AI
Tweet shares arXiv paper on diffusion-augmented LLMs for parallel token sampling.
AI
alphaXiv shares the post addressing fears that AI could make human thought unnecessary and proposes complementary intelligence.
